Why Single-Instruction Prompts Produce Generic Fiction
When you give an AI a single instruction — "write a sad scene," "write from a child's perspective," "write something mysterious" — you're essentially asking it to reach into its training data and find the statistical center of that concept. And the statistical center of "sad scene" is... every sad scene ever written, averaged together. You get the greatest hits of emotional clichés. You get the glistening eyes, the heavy silence, the words that catch in the throat.
AI models are, at their core, extraordinary pattern-matchers. Without additional constraints, they'll default to the most probable next word, the most expected phrase, the most well-worn emotional beat. This isn't a flaw — it's exactly what they're doing correctly. The model is giving you the most statistically "correct" version of what you asked for.
Generic outputs aren't a sign the AI is bad at writing. They're a sign you've given the AI too much freedom.
The paradox of creative AI: the more latitude you give it, the more ordinary the result. Constraints don't limit creativity — they redirect it toward specificity.
Think about how this works in your own writing. A blank page is terrifying precisely because anything is possible. But tell yourself "write a scene set in a laundromat, in second person, where nothing is said directly" and suddenly the creative problem becomes solvable. You're not less creative because of the constraints. You're more focused.
Constraint stacking does the same thing to an AI prompt — but with considerably more horsepower behind it.
What Constraint Stacking Is and How It Works
Constraint stacking is the practice of layering multiple simultaneous, specific instructions into a single prompt — not just about content, but about the formal and emotional properties of the writing itself. POV. Tense. Sentence rhythm. Emotional register. Vocabulary range. Structural patterns. You're not just telling the AI what to write about. You're telling it how the writing must feel, move, and land.
Each additional constraint narrows the probability space the model is working within. One constraint (sad scene) leaves a million possible outputs. Three constraints (sad scene, present tense, sentences under twelve words) leave maybe a thousand. Five or six well-chosen constraints? Now you're steering toward something genuinely specific.
Here's the mechanism in plain terms: every constraint you add is another filter on what counts as an acceptable output. The model can't just reach for the nearest cliché if the cliché violates your sentence-length rule, or breaks your POV instruction, or sits in the wrong emotional register. It has to find prose that satisfies all the conditions simultaneously — and that search almost always produces something more interesting than any single condition would alone.
This isn't about making prompts longer for the sake of it. A bloated prompt full of vague adjectives ("make it beautiful and evocative and powerful") is still a single-instruction prompt dressed up in more words. Real constraint stacking is about choosing constraints from different categories that create productive tension with each other.
The Four Constraint Categories Every Fiction Writer Should Know
Not all constraints are created equal. Telling an AI "write it more intensely" and telling it "write in close third person, present tense, with sentences that average five to seven words" are both constraints — but only one of them actually shapes the output in a predictable, controllable way. Over time, I've found that constraints cluster into four categories, and pulling from all four is where the magic happens.
1. Structural Constraints
These govern the formal architecture of the prose: POV (first, second, third limited, third omniscient), tense (past, present), sentence length and rhythm, paragraph length, and whether dialogue is present or absent. Structural constraints are the bones. They determine what the writing can and cannot do before a single word about emotion or character appears.
2. Linguistic Constraints
Vocabulary register (latinate vs. Anglo-Saxon, elevated vs. vernacular), the presence or absence of adverbs, punctuation patterns (em-dashes, fragments, semicolons), and reading level. A scene written with a directive to "use only one-syllable words wherever possible" reads completely differently than the same scene without that note — and often, more viscerally.
3. Emotional/Tonal Constraints
This is more than just "make it sad" or "make it funny." The granularity matters. Specify the emotional texture: grief laced with irritation, tenderness that the character is ashamed of, black humor as a coping mechanism. You can also specify what emotion to withhold — "the character is furious but the prose never shows it directly; it lives in what she notices."
4. Perceptual/Sensory Constraints
Which senses dominate? Is the character's attention primarily visual, auditory, kinesthetic? What do they notice and what do they ignore? A character who's a former chef will process a tense family dinner entirely differently than a character who's a musician. Pinning the perceptual lens forces the AI to write from inside a specific consciousness rather than from the outside looking in.
You don't need to use all four categories in every prompt. But the more categories you pull from, the further you push the output from generic and toward genuinely distinctive.
Building Your Stack: 4 Real Prompt Examples from Bland to Specific
Here's where this becomes practical. Below are four prompts moving from a single-instruction baseline up to a full constraint stack. Each one is something you could copy directly into a writing tool.
Prompt 1 — The Baseline (single instruction):
Write a scene where a woman realizes her marriage is over.You know what you'll get. She'll stare out a window. Her husband will say something mundane that suddenly sounds final. Her heart will sink. It's not bad — it's just the averaged output of ten thousand literary divorce scenes.
Prompt 2 — Adding structural constraints:
Write a scene where a woman realizes her marriage is over. Use close third person, present tense. Keep all sentences under ten words. No dialogue.Now the prose has to be percussive. Short sentences in present tense create a claustrophobic, breathless quality. The absence of dialogue forces the realization to live entirely in action and interiority. Already you're getting something more stylized. The constraint about sentence length is doing serious lifting here — it's nearly impossible to write a cliché smoothly in ten words or fewer.
Prompt 3 — Adding linguistic and emotional constraints:
Write a scene where a woman realizes her marriage is over. Use close third person, present tense. Keep all sentences under ten words. No dialogue. Use only simple Anglo-Saxon vocabulary — no latinate words. The emotional register is not grief but relief, which she finds shameful. She does not name this feeling. Show it only through what she chooses to do with her hands.That last detail — "what she chooses to do with her hands" — is doing something specific and powerful. It gives the AI a physical anchor for the emotion, which prevents the output from slipping into abstract internal monologue. The shame around relief is a much more complex emotional texture than straight sadness, and the prohibition on naming the feeling forces the AI to find a way to embody it rather than explain it. This prompt will produce something strange and specific almost every time.
Prompt 4 — Full constraint stack with perceptual layer:
Write a scene where a woman realizes her marriage is over. She's a former competitive swimmer — her sensory world is primarily physical, about water resistance, breath timing, the weight of her own body. Use close third person, present tense. Sentences under ten words. No dialogue. Simple Anglo-Saxon vocabulary only. The emotion is relief laced with shame; she never names it. She's doing the dishes. Her attention keeps moving to the temperature of the water on her hands and the specific sound the plates make. The word "love" cannot appear. Neither can "marriage" or "over."Notice what those final three prohibitions do. Banning the obvious words — "love," "marriage," "over" — forces the AI to find oblique ways to convey the scene's meaning. The swimmer's perceptual frame makes her physical sensations narratively legible; we understand something about how she processes experience before a single event occurs. The result will be strange, compressed, and genuinely unlike anything a single-instruction prompt could produce.
Tweak these by swapping in your character's profession, your story's tense, your specific forbidden words. The structure holds across genres and scenarios.
Common Mistakes: When Too Many Constraints Collapse the Output
Here's the thing nobody tells you: there's a breaking point. Stack too many constraints — especially contradictory ones — and the output doesn't become more distinctive. It becomes garbled, stiff, or incoherent. The AI is trying to satisfy incompatible conditions and either freezes into awkward prose or subtly ignores whichever constraints it finds hardest to honor.
The most common failure mode is contradictory structural constraints. Asking for "lyrical, flowing sentences" while simultaneously requiring "sentences under six words" is a direct contradiction. The AI will pick one and quietly abandon the other, usually the one you cared about more. Same with asking for "a stream-of-consciousness interiority" while also specifying "third person omniscient with narrative distance." These pull in opposite directions and produce muddled output.
The second failure mode is over-specifying content while under-specifying form. If you spend twelve lines describing the plot details of your scene but only add "make it literary" at the end, you've given the AI a content-heavy, form-light prompt. The content constraints will dominate, and you'll get a competent summary-adjacent summary of what you asked for. Flip the ratio: give less plot, more formal instruction.
The third failure is using emotional constraints that are too abstract. "Make it haunting" is not a constraint — it's a vibe. "The character is terrified but describes everything in cheerful, domestic terms" is a constraint. "There is something wrong but the prose never says so directly; it's in the rhythm of the sentences, which keep starting normally and then trailing off" is a constraint. The more behavioral and formal your emotional instructions, the more reliably the output will reflect them.
A practical rule: when an output comes back wrong, identify which constraint got dropped. AI models will consistently sacrifice certain types of constraints to honor others. Sentence length tends to get abandoned first when emotional content is very specific. POV tends to slip when tense is unusual (present tense in third person is genuinely less common in training data). Once you know what your model tends to sacrifice, you can either simplify the stack or reinforce the constraint you care most about at the end of the prompt, which gives it more weight.
If the output is stiff and mechanical, you have too many constraints. If it's smooth but generic, you have too few. The sweet spot is usually four to six specific, compatible constraints pulled from at least three categories.
Also worth noting: constraint stacking works best for scenes of a few hundred words. Asking for a full chapter under a complex constraint stack usually produces early excellence that degrades as the AI tries to maintain six rules across two thousand words. Use the stack to generate a distinctive passage, then use that passage as a style reference for longer generation. That's actually one of the most powerful workflows — generate a single constrained paragraph, then say "continue this scene, maintaining the same sentence rhythm and perceptual register," and let the established style carry forward.
The next time you get a flat, generic output, don't rewrite the scene description. Rewrite the formal instructions. Pick one structural constraint you haven't used, one linguistic rule, and one specific emotional texture with a behavioral expression. Drop your word count expectation, run it short, and see what comes back. That triangulation — structure, language, emotion, all specific, all compatible — is where AI stops producing prose that sounds like AI and starts producing prose that sounds like a choice someone made.